ANTsX evaluation in the BraTS-Reg22 challenge

• Training phase leaderboard
• Validation phase leaderboard
• Singularity container

Description

BraTS-Reg22 data were processed using previously vetted and frequently used registration parameter sets [1-2] which have been packaged within the different ANTsX platforms, specifically ANTs, ANTsPy, and ANTsR [3]. Each of these parameter sets consists of multiple transform stages for determining anatomical correspondence. Initial linear estimation of transform parameters comprises center of (intensity) mass alignment followed by optimization of both rigid and affine transforms using mutual information as the similarity metric [4]. The final deformable alignment utilizes symmetric normalization (SyN) with Gaussian [5] or B-spline [6] regularization and a neighborhood cross-correlation [5] or mutual information similarity metric. Also explored were the effects of image modality choice including all single modalities and combinations of modality pairs (specifically, T1-contrast enhanced/T2 and T1-contrast enhanced/FLAIR). Although performance with the training data was similar across the different configurations, SyN with Gaussian regularization and neighborhood cross-correlation (radius = 2 voxels) using T1-contrast enhanced images was selected for a single submission during the validation phase. Further details on this internal evaluation, including the precise ANTsPy calls, can be found at a dedicated GitHub repository [7].

1. Avants et al., A reproducible evaluation of ANTs similarity metric performance in brain image registration.
2. Avants et al., The Insight ToolKit Image Registration Framework.
3. https://github.com/ANTsX
4. Viola and Wells, Alignment by Maximization of Mutual Information.
5. Avants et al., Symmetric diffeomorphic image registration with cross-correlation: evaluating automated labeling of elderly and neurodegenerative brain.
6. Tustison and Avants, Explicit B-spline regularization in diffeomorphic image registration.
7. https://github.com/ntustison/BraTS-Reg22

Modalities to test on training data

• T1
• T1-contrast enhanced
• T2
• FLAIR
• T1-contrast enhanced + T2
• T1-contrast enhanced + FLAIR

Parameter sets to test on training data

• antsRegistrationSyNQuick[a]
  • Rigid stage
    • similarity metric: MI with 32 bins
    • gradient step = 0.1
    • iterations per level = [1000,500,250,0]
    • shrink factors per level = [8,4,2,1]
    • smoothing factor per level = [3,2,1,0] (voxels)
  • Affine stage
    • same parameters as affine stage
• antsRegistrationSyNQuick[s,32]
  • Rigid stage
    • Same as "antsRegistrationSyNQuick[a]"
  • Affine stage
    • Same as "antsRegistrationSyNQuick[a]"
  • SyN stage
    • Gaussian-based
    • similarity metric: MI with 32 bins
    • gradient step = 0.1
    • iterations per level = [100,70,50,0]
    • shrink factors per level = [8,4,2,1]
    • smoothing factor per level = [3,2,1,0] (voxels)
• antsRegistrationSyNQuick[b,32]
  • B-spline-based (mesh size per stage = [26,13,6.5,3.25] (mm))
  • other metrics same as "antsRegistrationSyNQuick[s,32]"
• antsRegistrationSyN[s,2]
  • Rigid stage
    • similarity metric: MI with 32 bins
    • gradient step = 0.1
    • iterations per level = [1000,500,250,100]
    • shrink factors per level = [8,4,2,1]
    • smoothing factor per level = [3,2,1,0] (voxels)
  • Affine stage
    • Same as rigid stage
  • SyN stage
    • Gaussian-based
    • similarity metric: CC with neighborhood radius of 2
    • gradient step = 0.1
    • iterations per level = [100,70,50,20]
    • shrink factors per level = [8,4,2,1]
    • smoothing factor per level = [3,2,1,0] (voxels)
• antsRegistrationSyN[b,2,26]
  • B-spline-based version of "antsRegistrationSyN[s,2]"
• antsRegistrationSyN[s,4]
  • Same as "antsRegistrationSyN[s,2]" with CC radius of 4
• antsRegistrationSyN[b,4,26]
  • B-spline-based version of "antsRegistrationSyN[s,4]"

Training data results

Accuracy

Although the organizers employed several metrics for evaluating performance, we employed the following "accuracy improvement" measure for assessment:

$$ \text{Accuracy improvement} = 1.0 - \frac{LandmarkError_{post}}{LandmarkError_{pre}} $$

Timing (in seconds)

• Mac Pro 2020
• 2.5 GHz 28-Core Intel Xeon W
• os.environ["ITK_GLOBAL_DEFAULT_NUMBER_OF_THREADS"] = "56"

Parameter set	Single modality	Two modalities
antsRegistrationSyNQuick[a]	4.6	NA
antsRegistrationSyNQuick[b,32,26]	39.8	38.5
antsRegistrationSyNQuick[s,32]	27.7	28.3
antsRegistrationSyN[s,2]	169.2	272.1
antsRegistrationSyN[b,2,26]	273.1	382.0
antsRegistrationSyN[s,4]	344.2	610.7
antsRegistrationSyN[b,4,26]	446.2	720.5

---

Apply antsRegistrationSyN[s,2] with T1-contrast enhanced to validation data

The choice of antsRegistrationSyN[s,2] with T1-contrast enhanced is due to combined performance and timing considerations.

>>> import ants
>>> import pandas as pd
>>>
>>> #######################################################################
>>> # I am making assumptions about file names.  Please change accordingly.
>>> 
>>> baseline_image = ants.image_read("baseline_t1ce.nii.gz")
>>> followup_image = ants.image_read("followup_t1ce.nii.gz")
>>> followup_landmark_file = "followup_landmarks.csv"
>>> warped_followup_landmark_file = "warped_followup_landmarks.csv"
>>>
>>> ########################################################################
>>> 
>>> reg = ants.registration(baseline_image, followup_image, type_of_transform="antsRegistrationSyN[s,2]", verbose=1)
>>>
>>> moving_indices = pd.read_csv(followup_landmark_file).drop('Landmark', axis=1)
>>> moving_indices = moving_indices.rename(columns={'X' : 'x', 'Y' : 'y', 'Z' : 'z'})
>>> moving_indices['y'] = moving_indices['y'] + 239
>>>
>>> moving_points = np.zeros(moving_indices.shape)
>>> for j in range(moving_indices.shape[0]):
>>>     moving_points[j,:] = ants.transform_index_to_physical_point(baseline_image, (moving_indices.iloc[j].values).astype(int))
>>> 
>>> moving_points_df = pd.DataFrame(data = {'x': moving_points[:,0], 'y': moving_points[:,1], 'z': moving_points[:,2]})
>>> moving_warped_points = ants.apply_transforms_to_points(3, moving_points_df, reg['invtransforms'], whichtoinvert=(True, False))   
>>> moving_warped_points = moving_warped_points.to_numpy()   
>>> moving_warped_points_df = pd.DataFrame(data=moving_warped_points, columns=['X', 'Y', 'Z'])
>>> moving_warped_points_df.insert(0, "Landmark", list(range(1, moving_points.shape[0]+1)))
>>> moving_warped_points_df.to_csv(warped_followup_landmark_file, index=False)

---

GoodhartsLaw